Re: [PATCH 5.17 127/298] driver core: Fix wait_for_device_probe() & deferred_probe_timeout interaction

From: Shreeya Patel
Date: Tue Aug 22 2023 - 10:12:21 EST

Hi Robin,

On 21/08/23 18:09, Robin Murphy wrote:
On 2023-08-21 12:35, Shreeya Patel wrote:

On 19/08/23 01:49, Saravana Kannan wrote:
On Thu, Aug 17, 2023 at 4:13 PM Shreeya Patel
<shreeya.patel@xxxxxxxxxxxxx> wrote:
Hi Geert, Saravana,

On 18/08/23 00:03, Saravana Kannan wrote:
On Thu, Aug 17, 2023 at 4:37 AM Shreeya Patel
<shreeya.patel@xxxxxxxxxxxxx> wrote:
Hi Greg,

On 16/08/23 20:33, Greg Kroah-Hartman wrote:
On Wed, Aug 16, 2023 at 03:09:27PM +0530, Shreeya Patel wrote:
On 13/06/22 15:40, Greg Kroah-Hartman wrote:
From: Saravana Kannan<saravanak@xxxxxxxxxx>

[ Upstream commit 5ee76c256e928455212ab759c51d198fedbe7523 ]

Mounting NFS rootfs was timing out when deferred_probe_timeout was
non-zero [1].  This was because ip_auto_config() initcall times out
waiting for the network interfaces to show up when
deferred_probe_timeout was non-zero. While ip_auto_config() calls
wait_for_device_probe() to make sure any currently running deferred
probe work or asynchronous probe finishes, that wasn't sufficient to
account for devices being deferred until deferred_probe_timeout.

Commit 35a672363ab3 ("driver core: Ensure wait_for_device_probe() waits
until the deferred_probe_timeout fires") tried to fix that by making
sure wait_for_device_probe() waits for deferred_probe_timeout to expire
before returning.

However, if wait_for_device_probe() is called from the kernel_init()
context:

- Before deferred_probe_initcall() [2], it causes the boot process to
      hang due to a deadlock.

- After deferred_probe_initcall() [3], it blocks kernel_init() from
      continuing till deferred_probe_timeout expires and beats the point of
      deferred_probe_timeout that's trying to wait for userspace to load
      modules.

Neither of this is good. So revert the changes to
wait_for_device_probe().

[1] -https://lore.kernel.org/lkml/TYAPR01MB45443DF63B9EF29054F7C41FD8C60@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
[2] -https://lore.kernel.org/lkml/YowHNo4sBjr9ijZr@dev-arch.thelio-3990X/
[3] -https://lore.kernel.org/lkml/Yo3WvGnNk3LvLb7R@xxxxxxxxxxxxx/
Hi Saravana, Greg,


KernelCI found this patch causes the baseline.bootrr.deferred-probe-empty test to fail on r8a77960-ulcb,
see the following details for more information.

KernelCI dashboard link:
https://linux.kernelci.org/test/plan/id/64d2a6be8c1a8435e535b264/

Error messages from the logs :-

+ UUID=11236495_1.5.2.4.5
+ set +x
+ export 'PATH=/opt/bootrr/libexec/bootrr/helpers:/lava-11236495/1/../bin:/sbin:/usr/sbin:/bin:/usr/bin'
+ cd /opt/bootrr/libexec/bootrr
+ sh helpers/bootrr-auto
e6800000.ethernet
e6700000.dma-controller
e7300000.dma-controller
e7310000.dma-controller
ec700000.dma-controller
ec720000.dma-controller
fea20000.vsp
feb00000.display
fea28000.vsp
fea30000.vsp
fe9a0000.vsp
fe9af000.fcp
fea27000.fcp
fea2f000.fcp
fea37000.fcp
sound
ee100000.mmc
ee140000.mmc
ec500000.sound
/lava-11236495/1/../bin/lava-test-case
<8>[   17.476741] <LAVA_SIGNAL_TESTCASE TEST_CASE_ID=deferred-probe-empty RESULT=fail>

Test case failing :-
Baseline Bootrr deferred-probe-empty test -https://github.com/kernelci/bootrr/blob/main/helpers/bootrr-generic-tests

Regression Reproduced :-

Lava job after reverting the commit 5ee76c256e92
https://lava.collabora.dev/scheduler/job/11292890


Bisection report from KernelCI can be found at the bottom of the email.

Thanks,
Shreeya Patel

#regzbot introduced: 5ee76c256e92
#regzbot title: KernelCI: Multiple devices deferring on r8a77960-ulcb

---------------------------------------------------------------------------------------------------------------------------------------------------

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * **
* If you do send a fix, please include this trailer: *
* Reported-by: "kernelci.org bot" <bot@...> *
* *
* Hope this helps! *
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

stable-rc/linux-5.10.y bisection: baseline.bootrr.deferred-probe-empty on
r8a77960-ulcb
You are testing 5.10.y, yet the subject says 5.17?

Which is it here?
Sorry, I accidentally used the lore link for 5.17 while reporting this
issue,
but this test does fail on all the stable releases from 5.10 onwards.

stable 5.15 :-
https://linux.kernelci.org/test/case/id/64dd156a5ac58d0cf335b1ea/
mainline :-
https://linux.kernelci.org/test/case/id/64dc13d55cb51357a135b209/

Shreeya, can you try the patch Geert suggested and let us know if it
helps? If not, then I can try to take a closer look.
I tried to test the kernel with 9be4cbd09da8 but it didn't change the
result.
https://lava.collabora.dev/scheduler/job/11311615

Also, I am not sure if this can change things but just FYI, KernelCI
adds some kernel parameters when running these tests and one of the
parameter is deferred_probe_timeout=60.
Ah this is good to know.

You can check this in the definition details given in the Lava job. I
also tried to remove this parameter and rerun the test but again I got
the same result.
How long does the test wait after boot before checking for the
deferred devices list?


AFAIK, script for running the tests is immediately ran after the boot process is complete so there is no wait time.

Regardless of what the kernel is doing, it seems like a fundamentally dumb test to specifically ask deferred probe to wait for up to a minute then complain that it hasn't finished after 11 seconds :/

If anything, it seems plausible that the "regression" might actually be the correct behaviour, and it was wrong before. I can't manage to pull up a boot log for a pre-5.10 kernel since all the async stuff on the KernelCI dashboard always just times out for me with a helpful "Error while loading data from the server (error code: 0)", but what would be interesting is whether those devices on the list are expected to successfully probe anyway - the mainline log below also shows other stuff failing to probe and CPUs failing to come online, so it's clearly not a very happy platform to begin with.


Sorry about the dashboard issues you are facing, KernelCI team is working on a new dashboard which will fix all of these issues. But we need to wait for it be ready for some more time.


Your point makes sense and that is why we did a test to add some sleep time of 60-65 seconds before running the tests and it actually fixed the problem. There are no more deferred devices as you can see in the following job.
https://lava.collabora.dev/scheduler/job/11330931

This change, to add deferred_probe_timeout=60 as kernel parameter was recently added in KernelCI since there were number of devices failing to probe on different platforms, specifically for chromebooks.
Unfortunately, no one realized to change the start of the test time and hence these issues are seen now on different platforms.

Thanks for pointing it out, it will help us to eliminate quite many deferred probe test failures that we are seeing.

Saravana, thanks to you as well for asking the valid questions about when the test starts running.

Thanks,
Shreeya Patel


#regzbot resolve: Nothing is broken due to the patch, a fix is needed on the KernelCI side.

Robin.

I will try to add 9be4cbd09da8 to mainline kernel and see what results I
get.
Now I'm confused. What do you mean by mainline? Are you saying the tip
of tree of Linus's tree is also hitting this issue?


KernelCI runs tests on different kernel branches and trees, we also have this same test running on mainline tree.
Following is the link to the dashboard for it and as you can see, it does fail there too.


https://linux.kernelci.org/test/case/id/64dc13d55cb51357a135b209/


-Saravana